Monaural Sound Source Separation by Perceptually Weighted Non-Negative Matrix Factorization - Demonstrations

The following audio examples demonstrate the performance of a perceptually weighted non-negative matrix factorization algorithm in one-channel sound source separation.

The signals are generated by separating the original 7-second excerpts into components, the number of which was set by hand (from 2 to 8 components were used). The components were clustered to sound sources by hand. Naturally, manual clustering may not be used in the final application. These examples are used just to demonstrate the possible outcome of a separation system.

original signal separated component 1 separated component 2 separated component 3
zone_orig.wav zone_sep1.wav zone_sep2.wav zone_sep3.wav
baby_orig.wav baby_sep1.wav baby_sep2.wav
science_orig.wav science_sep1.wav science_sep2.wav
eminem_orig.wav eminem_sep1.wav eminem_sep2.wav
watermelon_orig.wav watermelon_sep1.wav watermelon_sep2.wav

Demonstrations main page

Last modified: Tue Aug 30 13:25:03 EEST 2005 - Tuomas Virtanen, tuomas.virtanen@tut.fi